Research map / Architectures and efficiency / Pruning and efficient tuning
Structured network pruning
Prunes filters and channels with budget-aware regularization and probabilistic or continuous criteria.
162 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Pruning and efficient tuning, keyword cluster “sparsity · compression · structured”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
43 papers in 2023–24 and 29 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 15.5% to 5.2%.
No longer reported by papers here: CIFAR-10 (accuracy, last in 2024); CIFAR-100 (accuracy, last in 2024); ImageNet (Top-5 accuracy, last in 2021); ImageNet, ResNet-50, 90% sparsity (top-1 accuracy, last in 2024); CIFAR-10 (Top-1 accuracy, last in 2022).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Pruning Convolutional Neural Networks for Resource Efficient Inference (ICLR 2017 · 1,321 citations)
- 2018 · NISP: Pruning Networks Using Neuron Importance Score Propagation (CVPR 2018 · 844 citations)
- 2019 · Importance Estimation for Neural Network Pruning (CVPR 2019 · 956 citations)
- 2020 · HRank: Filter Pruning Using High-Rank Feature Map (CVPR 2020 · 888 citations)
- 2021 · Network Pruning via Performance Maximization (CVPR 2021 · 124 citations)
- 2022 · Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · 118 citations)
- 2023 · DepGraph: Towards Any Structural Pruning (CVPR 2023 · 484 citations)
- 2024 · Spanning Training Progress: Temporal Dual-Depth Scoring (TDDS) for Enhanced Dataset Pruning (CVPR 2024 · 19 citations)
- 2025 · Numerical Pruning for Efficient Autoregressive Models (AAAI 2025 · 11 citations)
- 2026 · UNSEEN: Enhancing Dataset Pruning from a Generalization Perspective (AAAI 2026 · 2 citations)
Earliest papers here
- Pruning Convolutional Neural Networks for Resource Efficient Inference (ICLR 2017 · 1,321 citations)
- Pruning Filters for Efficient ConvNets (ICLR 2017 · 666 citations)
- Runtime Neural Pruning (NeurIPS 2017 · 355 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
CIFAR-10 · accuracy (higher is better)
10 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 93.18 | Variational Convolutional Neural Network Pruning (CVPR 2019 · 381 citations) |
| 2020 | 94.1 | Towards Efficient Model Compression via Learned Global Ranking (CVPR 2020 · 176 citations) |
| 2021 | 94.75 | Network Pruning via Performance Maximization (CVPR 2021 · 124 citations) |
| 2022 | 95.07 | Prior Gradient Mask Guided Pruning-Aware Fine-Tuning (AAAI 2022 · 38 citations) |
ImageNet · Top-1 accuracy (higher is better)
10 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 75.33 | Global Sparse Momentum SGD for Pruning Very Deep Neural Networks (NeurIPS 2019 · 136 citations) |
| 2020 | 78.72 | Neural Network Pruning With Residual-Connections and Limited-Data (CVPR 2020 · 128 citations) |
CIFAR-100 · accuracy (higher is better)
6 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 73.76 | Variational Convolutional Neural Network Pruning (CVPR 2019 · 381 citations) |
| 2020 | 76.62 | Operation-Aware Soft Channel Pruning using Differentiable Masks (ICML 2020 · 32 citations) |
| 2024 | 78.54 | Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch (CVPR 2024 · 15 citations) |
ImageNet · Top-5 accuracy (higher is better)
6 papers, 2019–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 92.47 | Global Sparse Momentum SGD for Pruning Very Deep Neural Networks (NeurIPS 2019 · 136 citations) |
| 2020 | 93 | Towards Efficient Model Compression via Learned Global Ranking (CVPR 2020 · 176 citations) |
| 2021 | 93.77 | Network Pruning via Performance Maximization (CVPR 2021 · 124 citations) |
ImageNet, ResNet-50, 90% sparsity · top-1 accuracy (higher is better)
5 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 64.9 | Progressive Skeletonization: Trimming more fat from a network at initialization (ICLR 2021 · 40 citations) |
| 2022 | 66.86 | Prospect Pruning: Finding Trainable Weights at Initialization using Meta-Gradients (ICLR 2022 · 10 citations) |
| 2023 | 75.43 | A Unified Framework for Soft Threshold Pruning (ICLR 2023 · 5 citations) |
CIFAR-10 · Top-1 accuracy (higher is better)
4 papers, 2019–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 93.9 | Global Sparse Momentum SGD for Pruning Very Deep Neural Networks (NeurIPS 2019 · 136 citations) |
| 2020 | 94.53 | HRank: Filter Pruning Using High-Rank Feature Map (CVPR 2020 · 888 citations) |
What has been tried
138 papers here report 1,440 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 10%, hurt 15%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add the outer depth to EL2N | 30% pruning · accuracy | 76.89 -> 77.47 [+ 0.58] | Helped | Spanning Training Progress: Temporal Dual-Depth Scoring (TDDS) for Enhanced Dataset Pruning (CVPR 2024 · 19 citations) |
| use bi-level optimization instead of iterative training | CIFAR-10 · accuracy | 93.55% -> 93.78% | Helped | BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · 18 citations) |
| Choose the higher-accuracy subnet before finetuning | ImageNet-1K validation · Acc1 (%) after finetuning | 71.2 -> 71.8 | Helped | UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 2024 · 17 citations) |
| Set the regularization coefficient λ too small | CIFAR-10, ResNet-56 · test accuracy | Hurt | Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch (CVPR 2024 · 15 citations) | |
| Use 128 compensation samples instead of 512 or 1024 | compensation experiments across pruning ratios · perplexity | No clear effect | Numerical Pruning for Efficient Autoregressive Models (AAAI 2025 · 11 citations) | |
| train the original AlexNet for 20 epochs instead of 1 epoch | MNIST · Top1-Acc | 0.9902 -> 0.9920 | Helped | Entropy Induced Pruning Framework for Convolutional Neural Networks (AAAI 2024 · 8 citations) |
| Add SD-based pre-training before pruning with CE and SD fine-tuning (Simple Combination-2) | ImageNet, 90% sparsity · top-1 accuracy | 66.2% vs. 66.4% | Helped | EPSD: Early Pruning with Self-Distillation for Efficient Model Compression (AAAI 2024 · 7 citations) |
| Use attention routing rather than dynamic routing with α-entmax and orthogonality | CIFAR10 · accuracy | 70.01 -> 86.84 | Helped | OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning (CVPR 2024 · 7 citations) |
| Add expectation rescaling to soft pruning | CIFAR-100 · accuracy (%) | 77.5 -> 78.1 | Helped | InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning (ICLR 2024 · 7 citations) |
| Use a larger input size, increasing from 128 toward 224 | CIFAR100 · test accuracy | Helped | Visual Prompting Upgrades Neural Network Sparsification: A Data-Model Perspective (AAAI 2025 · 4 citations) | |
| Add multi-granularity pruning (block pruning) to OCP+LLM | ImageNet · accuracy-latency tradeoff | Helped | MDP: Multidimensional Vision Model Pruning with Latency Constraint (CVPR 2025 · 4 citations) | |
| Add BWT to DEPrune | ImageNet, MobileNet-V2 · DW-conv speedup | 2.75× → 3.35× | Helped | DEPrune: Depth-wise Separable Convolution Pruning for Maximizing GPU Parallelism (NeurIPS 2024 · 4 citations) |
What reviewers will expect
Datasets and baselines named in the 17 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR10 · 5 of 17 papers
- CIFAR100 · 5 of 17 papers
- ImageNet · 4 of 17 papers
- ImageNet-1K · 4 of 17 papers
- Tiny-ImageNet · 3 of 17 papers
- WikiText2 · 2 of 17 papers
- MNIST · 2 of 17 papers
- C4 · 2 of 17 papers
Baselines
- SparseGPT · 5 of 17 papers
- LLM-Pruner · 3 of 17 papers
- Wanda · 3 of 17 papers
- SliceGPT · 2 of 17 papers
- Magnitude pruning · 2 of 17 papers
- CHITA · 2 of 17 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 72 papers; from OpenAlex and the papers' affiliations).
- Google · 4 papers
- Samsung · 3 papers
- Huawei · 3 papers
- Meta · 3 papers
- Yonsei University · 2 papers
- Korea Institute of Science and Technology · 2 papers
- Xidian University · 2 papers
- Northeastern University · 2 papers
- National University of Singapore · 2 papers
- Zhejiang University · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Sparse training from scratch (95 papers)
- Joint pruning and quantization (78 papers)
- Layer-wise pruning of LLMs (91 papers)
Elsewhere in the topic
- Low-rank compression and merging (226 papers)
- Adapter-based parameter-efficient tuning (149 papers)
- LoRA variants and theory (121 papers)
- Parameter-efficient transfer analysis (97 papers)
- Efficient ViT adaptation (62 papers)
- Lottery ticket discovery (42 papers)
- Pruning at initialization (24 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).